Projects

Dashboard Video Segmentation, Fall 2021

For Digital Video Processing final, I implemented a transformer model for image segmentation using raw camera footage of a car’s dashboard camera.
Abstract:

In this paper, we cover the motivation and approach for accurate image segmentation from dashboard video. We specifically focus on performance for the Comma10k dataset, an open source dataset with 10,000 images with five separate segmentation classes. A variety of semantic segmentation architectures would perform well on this dataset, and we discuss their differences. The performance for our model of choice is compared to other public scores cited on the Comma10k repo, including Comma10k baseline. We cover important training and augmentation steps useful for performance gains and ultimately show great performance with contemporary transformer model Trans2Seg.

Dashboard Video Speed Prediction, Fall 2020

For Digital Image Processing final, I researched various methods for speed regression from raw camera footage of a car’s dashboard camera. These methods span simple linear regression models to end-to-end deep learning frameworks.
Abstract:

In this paper, we cover several approaches for accurate speed prediction from dashboard video. We specifically focus our attention to providing a sound solution to the Comma AI speed challenge [1]. We ultimately show great performance with a simple approach involving sparse optical flow calculations and keypoint detection derived from classical computer vision algorithms. This performance is compared to results obtained by replacin the classical algorithms with a deep learning model. Additionally, we cover our attempts using more sophisticated deep learning approaches that span topics such as interest point detection, visual odometry, and dense optical flow.